Recalculate and resend MTU probes if they are too large for the system.

[tinc] / src / net_packet.c

diff --git a/src/net_packet.c b/src/net_packet.c
index 9bebca4..a7ddcdf 100644 (file)
--- a/src/net_packet.c
+++ b/src/net_packet.c
@@ -37,6 +37,8 @@
 #include "digest.h"
 #include "device.h"
 #include "ethernet.h"
 #include "digest.h"
 #include "device.h"
 #include "ethernet.h"

+#include "ipv4.h"
+#include "ipv6.h"

 #include "graph.h"
 #include "logger.h"
 #include "net.h"
 #include "graph.h"
 #include "logger.h"
 #include "net.h"


@@ -69,7 +71,7 @@ static void try_fix_mtu(node_t *n) {
        if(n->mtuprobes < 0)
                return;
 
        if(n->mtuprobes < 0)
                return;
 

-       if(n->mtuprobes == 90 || n->minmtu >= n->maxmtu) {
+       if(n->mtuprobes == 20 || n->minmtu >= n->maxmtu) {

                if(n->minmtu > n->maxmtu)
                        n->minmtu = n->maxmtu;
                else
                if(n->minmtu > n->maxmtu)
                        n->minmtu = n->maxmtu;
                else


@@ -141,10 +143,11 @@ static void udp_probe_h(node_t *n, vpn_packet_t *packet, length_t len) {
                        timeout_add(&n->udp_ping_timeout, &udp_probe_timeout_handler, n, &(struct timeval){udp_discovery_timeout, 0});
                }
 
                        timeout_add(&n->udp_ping_timeout, &udp_probe_timeout_handler, n, &(struct timeval){udp_discovery_timeout, 0});
                }
 

-               if(probelen >= n->maxmtu + 8) {
+               if(probelen >= n->maxmtu + 1) {

                        logger(DEBUG_TRAFFIC, LOG_INFO, "Increase in PMTU to %s (%s) detected, restarting PMTU discovery", n->name, n->hostname);
                        n->maxmtu = MTU;
                        logger(DEBUG_TRAFFIC, LOG_INFO, "Increase in PMTU to %s (%s) detected, restarting PMTU discovery", n->name, n->hostname);
                        n->maxmtu = MTU;

-                       n->mtuprobes = 30;
+                       /* Set mtuprobes to 1 so that try_mtu() doesn't reset maxmtu */
+                       n->mtuprobes = 1;

                        return;
                }
 
                        return;
                }
 


@@ -157,10 +160,9 @@ static void udp_probe_h(node_t *n, vpn_packet_t *packet, length_t len) {
                        try_fix_mtu(n);
                }
 
                        try_fix_mtu(n);
                }
 

-               /* Calculate RTT and bandwidth.
+               /* Calculate RTT.

                   The RTT is the time between the MTU probe burst was sent and the first
                   The RTT is the time between the MTU probe burst was sent and the first

-                  reply is received. The bandwidth is measured using the time between the
-                  arrival of the first and third probe reply (or type 2 probe requests).
+                  reply is received.

                 */
 
                struct timeval now, diff;
                 */
 
                struct timeval now, diff;


@@ -180,12 +182,7 @@ static void udp_probe_h(node_t *n, vpn_packet_t *packet, length_t len) {
                if(n->probe_counter == 1) {
                        n->rtt = diff.tv_sec + diff.tv_usec * 1e-6;
                        n->probe_time = probe_timestamp;
                if(n->probe_counter == 1) {
                        n->rtt = diff.tv_sec + diff.tv_usec * 1e-6;
                        n->probe_time = probe_timestamp;

-               } else if(n->probe_counter == 3) {
-                       /* TODO: this will never fire - we're not sending batches of three anymore. */
-                       struct timeval probe_timestamp_diff;
-                       timersub(&probe_timestamp, &n->probe_time, &probe_timestamp_diff);
-                       n->bandwidth = 2.0 * probelen / (probe_timestamp_diff.tv_sec + probe_timestamp_diff.tv_usec * 1e-6);
-                       logger(DEBUG_TRAFFIC, LOG_DEBUG, "%s (%s) RTT %.2f ms, burst bandwidth %.3f Mbit/s, rx packet loss %.2f %%", n->name, n->hostname, n->rtt * 1e3, n->bandwidth * 8e-6, n->packetloss * 1e2);
+                       logger(DEBUG_TRAFFIC, LOG_DEBUG, "%s (%s) RTT %.2f ms, rx packet loss %.2f %%", n->name, n->hostname, n->rtt * 1e3, n->packetloss * 1e2);

                }
        }
 }
                }
        }
 }


@@ -889,6 +886,67 @@ static void try_udp(node_t* n) {
        }
 }
 
        }
 }
 

+static length_t choose_initial_maxmtu(node_t *n) {
+#ifdef IP_MTU
+
+       int sock = -1;
+
+       const sockaddr_t *sa = NULL;
+       int sockindex;
+       choose_udp_address(n, &sa, &sockindex);
+       if(!sa)
+               return MTU;
+
+       sock = socket(sa->sa.sa_family, SOCK_DGRAM, IPPROTO_UDP);
+       if(sock < 0) {
+               logger(DEBUG_TRAFFIC, LOG_ERR, "Creating MTU assessment socket for %s (%s) failed: %s", n->name, n->hostname, sockstrerror(sockerrno));
+               return MTU;
+       }
+
+       if(connect(sock, &sa->sa, SALEN(sa->sa))) {
+               logger(DEBUG_TRAFFIC, LOG_ERR, "Connecting MTU assessment socket for %s (%s) failed: %s", n->name, n->hostname, sockstrerror(sockerrno));
+               close(sock);
+               return MTU;
+       }
+
+       int ip_mtu;
+       socklen_t ip_mtu_len = sizeof ip_mtu;
+       if(getsockopt(sock, IPPROTO_IP, IP_MTU, &ip_mtu, &ip_mtu_len)) {
+               logger(DEBUG_TRAFFIC, LOG_ERR, "getsockopt(IP_MTU) on %s (%s) failed: %s", n->name, n->hostname, sockstrerror(sockerrno));
+               close(sock);
+               return MTU;
+       }
+
+       close(sock);
+
+       /* getsockopt(IP_MTU) returns the MTU of the physical interface.
+          We need to remove various overheads to get to the tinc MTU. */
+       length_t mtu = ip_mtu;
+       mtu -= (sa->sa.sa_family == AF_INET6) ? sizeof(struct ip6_hdr) : sizeof(struct ip);
+       mtu -= 8; /* UDP */
+       if(n->status.sptps) {
+               mtu -= SPTPS_DATAGRAM_OVERHEAD;
+               if((n->options >> 24) >= 4)
+                       mtu -= sizeof(node_id_t) + sizeof(node_id_t);
+       }
+
+       if (mtu < 512) {
+               logger(DEBUG_TRAFFIC, LOG_ERR, "getsockopt(IP_MTU) on %s (%s) returned absurdly small value: %d", n->name, n->hostname, ip_mtu);
+               return MTU;
+       }
+       if (mtu > MTU)
+               return MTU;
+
+       logger(DEBUG_TRAFFIC, LOG_INFO, "Using system-provided maximum tinc MTU for %s (%s): %hd", n->name, n->hostname, mtu);
+       return mtu;
+
+#else
+
+       return MTU;
+
+#endif
+}
+

 // This function tries to determines the MTU of a node.
 // By calling this function repeatedly, n->minmtu will be progressively increased, and at some point, n->mtu will be fixed to n->minmtu.
 // If the MTU is already fixed, this function checks if it can be increased.
 // This function tries to determines the MTU of a node.
 // By calling this function repeatedly, n->minmtu will be progressively increased, and at some point, n->mtu will be fixed to n->minmtu.
 // If the MTU is already fixed, this function checks if it can be increased.


@@ -903,8 +961,8 @@ static void try_mtu(node_t *n) {
                return;
        }
 
                return;
        }
 

-       /* mtuprobes == 0..89: initial discovery, send bursts with 1 second interval, mtuprobes++
-          mtuprobes ==    90: fix MTU, and go to -1
+       /* mtuprobes == 0..19: initial discovery, send bursts with 1 second interval, mtuprobes++
+          mtuprobes ==    20: fix MTU, and go to -1

           mtuprobes ==    -1: send one >maxmtu probe every pingtimeout */
 
        struct timeval now;
           mtuprobes ==    -1: send one >maxmtu probe every pingtimeout */
 
        struct timeval now;


@@ -925,15 +983,48 @@ static void try_mtu(node_t *n) {
        if(n->mtuprobes < 0) {
                /* After the initial discovery, we only send one >maxmtu probe
                   to detect PMTU increases. */
        if(n->mtuprobes < 0) {
                /* After the initial discovery, we only send one >maxmtu probe
                   to detect PMTU increases. */

-               if(n->maxmtu + 8 < MTU)
-                       send_udp_probe_packet(n, n->maxmtu + 8);
+               if(n->maxmtu + 1 < MTU)
+                       send_udp_probe_packet(n, n->maxmtu + 1);

        } else {
        } else {

-               /* Probes are sent with random sizes between the
-                  lower and upper boundaries for the MTU thus far discovered. */
-               int len = n->maxmtu;
-               if(n->minmtu < n->maxmtu)
-                       len = n->minmtu + 1 + rand() % (n->maxmtu - n->minmtu);
-               send_udp_probe_packet(n, MAX(len, 64));
+               /* Before initial discovery begins, set maxmtu to the most likely value.
+                  If it's underestimated, we will correct it after initial discovery. */
+               if(n->mtuprobes == 0)
+                       n->maxmtu = choose_initial_maxmtu(n);
+
+               for (;;) {
+                       /* Decreasing the number of probes per cycle might make the algorithm react faster to lost packets,
+                          but it will typically increase convergence time in the no-loss case. */
+                       const length_t probes_per_cycle = 8;
+
+                       /* This magic value was determined using math simulations.
+                          It will result in a 1329-byte first probe, followed (if there was a reply) by a 1407-byte probe.
+                          Since 1407 is just below the range of tinc MTUs over typical networks,
+                          this fine-tuning allows tinc to cover a lot of ground very quickly.
+                          This fine-tuning is only valid for maxmtu = MTU; if maxmtu is smaller,
+                          then it's better to use a multiplier of 1. Indeed, this leads to an interesting scenario
+                          if choose_initial_maxmtu() returns the actual MTU value - it will get confirmed with one single probe. */
+                       const float multiplier = (n->maxmtu == MTU) ? 0.97 : 1;
+
+                       const float cycle_position = probes_per_cycle - (n->mtuprobes % probes_per_cycle) - 1;
+                       const length_t minmtu = MAX(n->minmtu, 512);
+                       const float interval = n->maxmtu - minmtu;
+
+                       /* The core of the discovery algorithm is this exponential.
+                          It produces very large probes early in the cycle, and then it very quickly decreases the probe size.
+                          This reflects the fact that in the most difficult cases, we don't get any feedback for probes that
+                          are too large, and therefore we need to concentrate on small offsets so that we can quickly converge
+                          on the precise MTU as we are approaching it.
+                          The last probe of the cycle is always 1 byte in size - this is to make sure we'll get at least one
+                          reply per cycle so that we can make progress. */
+                       const length_t offset = powf(interval, multiplier * cycle_position / (probes_per_cycle - 1));
+
+                       length_t maxmtu = n->maxmtu;
+                       send_udp_probe_packet(n, minmtu + offset);
+                       /* If maxmtu changed, it means the probe was rejected by the system because it was too large.
+                          In that case, we recalculate with the new maxmtu and try again. */
+                       if(n->mtuprobes < 0 || maxmtu == n->maxmtu)
+                               break;
+               }

 
                if(n->mtuprobes >= 0)
                        n->mtuprobes++;
 
                if(n->mtuprobes >= 0)
                        n->mtuprobes++;